Self-Supervised Learning for Anomaly Detection in IoT Networks

Due to the fast growth of the IoT, security challenges are rising notably in anomaly detection. Traditional supervised learning techniques heavily depend on labeled datasets, which are often rare and become outdated as cyber adversaries evolve. Unsupervised learning is very powerful but produces high false-positive rates and lacks context. To overcome these limitations, this study presents a self-supervised learning structure for detecting anomalies in IoT networks. Utilizing contrastive learning, autoencoders, and transformers, the model derives useful representations from unlabeled data, elevating detection accuracy whilst minimizing reliance on manual labels. With the TON_IoT dataset, our proposed model outperforms traditional supervised and unsupervised methods, achieving 0.94 accuracy, 0.93 precision, and an AUC-ROC of 0.96. The outcomes illustrate the ability of self-supervised learning to improve cybersecurity in dynamic IoT environments.